Why Is GPT-5.5 Disappointing?

TL;DR

GPT-5.5 disappoints despite its power mainly because of a steep price hike, no API support at launch, and a tendency to need explicit guidance to honor user intent. It costs $5 per million tokens in and $30 out, roughly 2x GPT-5.4 and about 20% above Opus 4.7. It is more token-efficient and crushes coding benchmarks (82.7% on terminal bench vs 75.1 before), and the Pro version impresses, but the rough edges make it hard to love. Here is how to get the most from it.

Transcript

Sup nerds. You might notice a couple things different about my setup because I'm currently in Miami at an event. But this is an important enough release I wanted to cover it and I think my coverage might surprise you guys a bit. If you couldn't tell from the title, this is a video about GPT 5.5, the latest model released by OpenAI. And I'm sure you... Read More

Key Insights

• GPT-5.5 is the latest model from OpenAI, offering improved token efficiency but at a significantly higher cost than its predecessors.
• The model is designed with enhanced safeguards to reduce misuse, incorporating feedback from early access partners.
• Despite its intelligence, GPT-5.5 sometimes struggles with honoring user intent, often requiring more explicit guidance.
• The model's lack of API support limits its flexibility for developers, although a workaround exists through the codeex endpoints.
• GPT-5.5 excels in coding tasks, outperforming previous models in benchmarks and real-world applications.
• The Pro version of GPT-5.5 demonstrates exceptional problem-solving abilities, though it is not yet available via API.
• Users may need to create new threads frequently due to the model's difficulty in clearing incorrect context information.
• The model's high cost is partly justified by its token efficiency, but users must adapt their approach to fully leverage its capabilities.

Questions & Answers

Q: Is GPT-5.5 lazy, or does it just need more guidance?

It can feel lazy because it often struggles to honor user intent and requires more explicit direction to produce what you want. In practice you need to craft detailed prompts and spell out expectations rather than relying on the model to infer them. Spending more time on prompt preparation is the main way to get reliable results.

Q: How much does GPT-5.5 cost compared to previous models?

GPT-5.5 costs $5 per million tokens in and $30 per million tokens out. That is 2x the price of GPT-5.4 and around 20% higher than Opus 4.7. The model is much more token-efficient than either, which partly offsets the hike, but the increase is still large.

Q: Does GPT-5.5 have API support?

Not at launch. API support does not exist yet, though OpenAI says it will arrive soon and workarounds exist through the codex endpoints. For now the model is mainly usable through the ChatGPT site, which limits what developers can build with it.

Q: How does GPT-5.5 perform on coding benchmarks?

It leads on coding. On terminal bench it scored 82.7% versus 75.1 for the prior model, and on OpenAI's internal expert SWE bench it hit 73% versus 68.5 before. It writes more efficient, reliable code and outperforms earlier models, though results still improve when you give it explicit guidance.

Q: What is the GPT-5.5 Pro version like?

The Pro version stands out for problem-solving and is the most impressive part of the release. OpenAI does not run it on standard benchmarks because it is not meant for harnesses, and it is only accessible through the ChatGPT site rather than the API. For advanced, complex tasks it is the strongest option here.

Q: Why do long-running threads cause problems in GPT-5.5?

The model tends to retain incorrect context and has difficulty clearing it, so mistakes can repeat across a long conversation. The recommended fix is to create new threads frequently rather than continuing a polluted one. Combined with detailed prompts, this keeps the model on track.

Q: What safeguards does GPT-5.5 ship with?

OpenAI calls it their strongest set of safeguards to date, designed to reduce misuse while preserving beneficial access. They used internal and external red teamers, added targeted testing for advanced cybersecurity and biology capabilities, and collected feedback from nearly 200 trusted early access partners before release.

Q: What hardware is GPT-5.5 built on?

It is a larger model that is normally harder to serve at speed, but OpenAI made changes to address that, largely through a partnership with Nvidia. It was trained and shipped on Nvidia's latest GB200 NVL72 systems, and Nvidia has been happy with the results.

Summary & Key Takeaways

• GPT-5.5 is a powerful AI model with improved token efficiency but comes with a steep price increase. Users may find it challenging to manage long-running threads due to the model's tendency to retain incorrect context information. Despite its coding prowess, the lack of API support limits its flexibility.

• The model is equipped with advanced safeguards to prevent misuse, and it excels in coding tasks, outperforming previous models in benchmarks. However, users must provide explicit guidance to achieve desired results, as the model sometimes struggles to honor user intent.

• GPT-5.5's Pro version showcases exceptional problem-solving abilities, though it is only accessible through the ChatGPT site. Users are advised to craft detailed prompts and create new threads frequently to maximize the model's potential, despite its high cost.